afk

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

AFK

The user hands over a task and leaves. Settle every question up front, then work alone until the chosen end state or a stop.

用户交付任务后离开。事先解决所有问题，然后在无人干预的情况下工作，直到达到选定的终态或触发停止条件。

1. Launch

1. 启动

The user waits on nothing. In the first reply:

Print this question before any tool call:

markdown

1. End with?
   1.1 staged changes
   1.2 open PR
   1.3 merged PR

Spawn a researcher agent in the background. Pick its model per the repo's model rule. Task: read the issue with its comments, the code it names,
```
CONTEXT.md
```
and related ADRs, then return a draft of the brief below.
End the turn.

Numbering follows

.claude/rules/workflow-dialog.md

用户无需等待任何东西。在首次回复中：

在任何工具调用之前打印此问题：

markdown

1. End with?
   1.1 staged changes
   1.2 open PR
   1.3 merged PR

在后台启动 researcher Agent。根据仓库的模型规则选择模型。任务：阅读 issue 及其评论、issue 中提到的代码、
```
CONTEXT.md
```
和相关 ADR，然后返回下方 brief 的草稿。
结束本轮。

编号方式遵循

.claude/rules/workflow-dialog.md

。

2. Brief

When the researcher returns, send one message, whether or not question 1 has an answer yet. An open question 1 stays in the message, so the user answers everything in one reply.

What I'd do: at most three bullets. Name the files and areas you plan to touch. This is the scope.
Questions: only what blocks the work or forks the solution. Numbered, one sub-item per option, recommended marks your pick. No questions: omit the section.
Last line:
```
go?
```

Repeat until the user approves. A changed scope or end state needs a new

go?

当 researcher 返回时，发送一条消息，无论问题 1 是否已经得到回答。尚未回答的问题 1 仍需保留在消息中，以便用户一次性回复所有问题。

**What I'd do：**最多三个要点。列出计划修改的文件和区域。这就是工作范围。
**Questions：**仅包含会阻塞工作或导致方案分歧的问题。使用编号编号，每个选项使用一个子项，recommended 标记你的选择。如果没有疑问，则省略此部分。
最后一行：
```
go?
```

重复此过程，直到用户批准。范围或终态发生变化时，需要重新获得一次

go?

。

3. Go

3. 开始

On approval, print one line and start work:

starting now... <animal> _<sound>_ "<joke>"

Pick the animal from this list that fits the task best: 🐒🦍🦧🐕🐩🐺🦊🦝🐈🐈‍⬛🦁🐯🐅🐆🫎🫏🐎🦄🦓🦌🦬🐂🐃🐄🐖🐗🐏🐑🐐🐫🦙🦒🐘🦣🦏🦛🐁🐀🐇🐿️🦫🦔🦇🐻🐨🐼🦥🦦🦨🦘🦡🦃🐓🐥🐧🕊️🦅🦆🦢🦉🦤🦩🦚🦜🐦‍⬛🪿🐦‍🔥🐸🐊🐢🦎🐍🐉🦕🦖🐳🐋🐬🦭🐠🐡🦈🐙🪼🦀🦑🐌🦋🐛🐜🐝🪲🐞🦗🪳🕷️🦂🦟🪰🪱🦠
Sound: the noise that animal makes, italic.
Joke: one line, about both the animal and the task.

Example, task "remove the old login auth system":

starting now... 🦖 _Roaar_ "Why did the T-Rex die during the meteor hit? He still had the old login credentials for the underground bunker."

获得批准后，打印一行并开始工作：

starting now... <animal> _<sound>_ "<joke>"

从列表中选择最适合该任务的动物：🐒🦍🦧🐕🐩🐺🦊🦝🐈🐈‍⬛🦁🐯🐅🐆🫎🫏🐎🦄🦓🦌🦬🐂🐃🐄🐖🐗🐏🐑🐐🐫🦙🦒🐘🦣🦏🦛🐁🐀🐇🐿️🦫🦔🦇🐻🐨🐼🦥🦦🦨🦘🦡🦃🐓🐥🐧🕊️🦅🦆🦢🦉🦤🦩🦚🦜🐦‍⬛🪿🐦‍🔥🐸🐊🐢🦎🐍🐉🦕🦖🐳🐋🐬🦭🐠🐡🦈🐙🪼🦀🦑🐌🦋🐛🐜🐝🪲🐞🦗🪳🕷️🦂🦟🪰🪱🦠
声音：使用该动物发出的声音，并以斜体显示。
笑话：一行内容，需要同时与动物和任务相关。

示例，任务为“移除旧的登录认证系统”：

starting now... 🦖 _Roaar_ "Why did the T-Rex die during the meteor hit? He still had the old login credentials for the underground bunker."

4. Work

4. 工作

Build, test, verify per the repo's rules. Ask nothing between

go

and the end. A question means a stop.

Stop when:

Confidence drops. You doubt the approach works, you found a better approach that needs a major pivot, or you guess at intent the brief did not settle.
Blast radius grows. The change needs a file or area outside the scope, a new public API, a schema change or a new dependency.
Blocked. Missing access, secret or tool, or a decision only the user can make.
Red loop. The same check stays red after 3 fix attempts.

On a stop, leave the work uncommitted and print the report in the

/progress

shape (

.agents/skills/progress/SKILL.md

), the stop reason under Needs you.

按照仓库规则进行构建、测试和验证。在

go

与结束之间不要提出任何问题。提出问题即意味着停止。

出现以下情况时停止：

**信心下降。**你怀疑当前方案是否有效，发现了需要重大调整的更优方案，或者在猜测 brief 尚未明确的意图。
**影响范围扩大。**变更需要修改范围之外的文件或区域，需要新增公共 API、进行 schema 变更或引入新依赖。
**受到阻塞。**缺少访问权限、secret 或工具，或者需要用户才能做出的决策。
**Red loop。**同一项检查在尝试修复 3 次后仍然为红灯。

停止时，保持工作内容未提交，并按照

/progress

格式（

.agents/skills/progress/SKILL.md

）打印报告，在 Needs you 下说明停止原因。

5. End

5. 结束

The end state answer is the user's sign-off in advance. It replaces the approval step of

.claude/rules/workflow-sign-off.md

for this task only.

End state	Do
staged changes	Stage the work, stop.
open PR	Commit, push, open the PR. Stop at green CI.
merged PR	As open PR, then apply `.agents/skills/babysit-pr/SKILL.md` with `--merge` .

CI failures and review items after the push count toward the stop rules above.

Close with the report in the

/progress

shape: end state reached, links (PR, commits), what was verified and how.

终态选项代表用户事先给出的签核。对于本任务，它取代

.claude/rules/workflow-sign-off.md

中的批准步骤。

终态	操作
staged changes	暂存工作内容，然后停止。
open PR	提交、推送并开启 PR。CI 变为绿灯后停止。
merged PR	按照 open PR 的流程操作，然后使用 `--merge` 应用 `.agents/skills/babysit-pr/SKILL.md` 。

推送后出现的 CI 失败和审查意见，也计入上述停止规则。

最后按照

/progress

格式关闭并输出报告：说明已达成终态，提供相关链接（PR、commits），并列出验证内容及验证方式。